drivers/md/bcache/request.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Main bcache entry point - handle a read or a write request and decide what to
   4  * do with it; the make_request functions are called by the block layer.
   5  *
   6  * Copyright 2010, 2011 Kent Overstreet <[email protected]>
   7  * Copyright 2012 Google, Inc.
   8  */
   9
  10 #include "bcache.h"
  11 #include "btree.h"
  12 #include "debug.h"
  13 #include "request.h"
  14 #include "writeback.h"
  15
  16 #include <linux/module.h>
  17 #include <linux/hash.h>
  18 #include <linux/random.h>
  19 #include <linux/backing-dev.h>
  20
  21 #include <trace/events/bcache.h>
  22
  23 #define CUTOFF_CACHE_ADD        95
  24 #define CUTOFF_CACHE_READA      90
  25
  26 struct kmem_cache *bch_search_cache;
  27
  28 static void bch_data_insert_start(struct closure *);
  29
  30 static unsigned cache_mode(struct cached_dev *dc)
  31 {
  32         return BDEV_CACHE_MODE(&dc->sb);
  33 }
  34
  35 static bool verify(struct cached_dev *dc)
  36 {
  37         return dc->verify;
  38 }
  39
  40 static void bio_csum(struct bio *bio, struct bkey *k)
  41 {
  42         struct bio_vec bv;
  43         struct bvec_iter iter;
  44         uint64_t csum = 0;
  45
  46         bio_for_each_segment(bv, bio, iter) {
  47                 void *d = kmap(bv.bv_page) + bv.bv_offset;
  48                 csum = bch_crc64_update(csum, d, bv.bv_len);
  49                 kunmap(bv.bv_page);
  50         }
  51
  52         k->ptr[KEY_PTRS(k)] = csum & (~0ULL >> 1);
  53 }
  54
  55 /* Insert data into cache */
  56
  57 static void bch_data_insert_keys(struct closure *cl)
  58 {
  59         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
  60         atomic_t *journal_ref = NULL;
  61         struct bkey *replace_key = op->replace ? &op->replace_key : NULL;
  62         int ret;
  63
  64         /*
  65          * If we're looping, might already be waiting on
  66          * another journal write - can't wait on more than one journal write at
  67          * a time
  68          *
  69          * XXX: this looks wrong
  70          */
  71 #if 0
  72         while (atomic_read(&s->cl.remaining) & CLOSURE_WAITING)
  73                 closure_sync(&s->cl);
  74 #endif
  75
  76         if (!op->replace)
  77                 journal_ref = bch_journal(op->c, &op->insert_keys,
  78                                           op->flush_journal ? cl : NULL);
  79
  80         ret = bch_btree_insert(op->c, &op->insert_keys,
  81                                journal_ref, replace_key);
  82         if (ret == -ESRCH) {
  83                 op->replace_collision = true;
  84         } else if (ret) {
  85                 op->status              = BLK_STS_RESOURCE;
  86                 op->insert_data_done    = true;
  87         }
  88
  89         if (journal_ref)
  90                 atomic_dec_bug(journal_ref);
  91
  92         if (!op->insert_data_done) {
  93                 continue_at(cl, bch_data_insert_start, op->wq);
  94                 return;
  95         }
  96
  97         bch_keylist_free(&op->insert_keys);
  98         closure_return(cl);
  99 }
 100
 101 static int bch_keylist_realloc(struct keylist *l, unsigned u64s,
 102                                struct cache_set *c)
 103 {
 104         size_t oldsize = bch_keylist_nkeys(l);
 105         size_t newsize = oldsize + u64s;
 106
 107         /*
 108          * The journalling code doesn't handle the case where the keys to insert
 109          * is bigger than an empty write: If we just return -ENOMEM here,
 110          * bio_insert() and bio_invalidate() will insert the keys created so far
 111          * and finish the rest when the keylist is empty.
 112          */
 113         if (newsize * sizeof(uint64_t) > block_bytes(c) - sizeof(struct jset))
 114                 return -ENOMEM;
 115
 116         return __bch_keylist_realloc(l, u64s);
 117 }
 118
 119 static void bch_data_invalidate(struct closure *cl)
 120 {
 121         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 122         struct bio *bio = op->bio;
 123
 124         pr_debug("invalidating %i sectors from %llu",
 125                  bio_sectors(bio), (uint64_t) bio->bi_iter.bi_sector);
 126
 127         while (bio_sectors(bio)) {
 128                 unsigned sectors = min(bio_sectors(bio),
 129                                        1U << (KEY_SIZE_BITS - 1));
 130
 131                 if (bch_keylist_realloc(&op->insert_keys, 2, op->c))
 132                         goto out;
 133
 134                 bio->bi_iter.bi_sector  += sectors;
 135                 bio->bi_iter.bi_size    -= sectors << 9;
 136
 137                 bch_keylist_add(&op->insert_keys,
 138                                 &KEY(op->inode, bio->bi_iter.bi_sector, sectors));
 139         }
 140
 141         op->insert_data_done = true;
 142         /* get in bch_data_insert() */
 143         bio_put(bio);
 144 out:
 145         continue_at(cl, bch_data_insert_keys, op->wq);
 146 }
 147
 148 static void bch_data_insert_error(struct closure *cl)
 149 {
 150         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 151
 152         /*
 153          * Our data write just errored, which means we've got a bunch of keys to
 154          * insert that point to data that wasn't succesfully written.
 155          *
 156          * We don't have to insert those keys but we still have to invalidate
 157          * that region of the cache - so, if we just strip off all the pointers
 158          * from the keys we'll accomplish just that.
 159          */
 160
 161         struct bkey *src = op->insert_keys.keys, *dst = op->insert_keys.keys;
 162
 163         while (src != op->insert_keys.top) {
 164                 struct bkey *n = bkey_next(src);
 165
 166                 SET_KEY_PTRS(src, 0);
 167                 memmove(dst, src, bkey_bytes(src));
 168
 169                 dst = bkey_next(dst);
 170                 src = n;
 171         }
 172
 173         op->insert_keys.top = dst;
 174
 175         bch_data_insert_keys(cl);
 176 }
 177
 178 static void bch_data_insert_endio(struct bio *bio)
 179 {
 180         struct closure *cl = bio->bi_private;
 181         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 182
 183         if (bio->bi_status) {
 184                 /* TODO: We could try to recover from this. */
 185                 if (op->writeback)
 186                         op->status = bio->bi_status;
 187                 else if (!op->replace)
 188                         set_closure_fn(cl, bch_data_insert_error, op->wq);
 189                 else
 190                         set_closure_fn(cl, NULL, NULL);
 191         }
 192
 193         bch_bbio_endio(op->c, bio, bio->bi_status, "writing data to cache");
 194 }
 195
 196 static void bch_data_insert_start(struct closure *cl)
 197 {
 198         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 199         struct bio *bio = op->bio, *n;
 200
 201         if (op->bypass)
 202                 return bch_data_invalidate(cl);
 203
 204         if (atomic_sub_return(bio_sectors(bio), &op->c->sectors_to_gc) < 0)
 205                 wake_up_gc(op->c);
 206
 207         /*
 208          * Journal writes are marked REQ_PREFLUSH; if the original write was a
 209          * flush, it'll wait on the journal write.
 210          */
 211         bio->bi_opf &= ~(REQ_PREFLUSH|REQ_FUA);
 212
 213         do {
 214                 unsigned i;
 215                 struct bkey *k;
 216                 struct bio_set *split = &op->c->bio_split;
 217
 218                 /* 1 for the device pointer and 1 for the chksum */
 219                 if (bch_keylist_realloc(&op->insert_keys,
 220                                         3 + (op->csum ? 1 : 0),
 221                                         op->c)) {
 222                         continue_at(cl, bch_data_insert_keys, op->wq);
 223                         return;
 224                 }
 225
 226                 k = op->insert_keys.top;
 227                 bkey_init(k);
 228                 SET_KEY_INODE(k, op->inode);
 229                 SET_KEY_OFFSET(k, bio->bi_iter.bi_sector);
 230
 231                 if (!bch_alloc_sectors(op->c, k, bio_sectors(bio),
 232                                        op->write_point, op->write_prio,
 233                                        op->writeback))
 234                         goto err;
 235
 236                 n = bio_next_split(bio, KEY_SIZE(k), GFP_NOIO, split);
 237
 238                 n->bi_end_io    = bch_data_insert_endio;
 239                 n->bi_private   = cl;
 240
 241                 if (op->writeback) {
 242                         SET_KEY_DIRTY(k, true);
 243
 244                         for (i = 0; i < KEY_PTRS(k); i++)
 245                                 SET_GC_MARK(PTR_BUCKET(op->c, k, i),
 246                                             GC_MARK_DIRTY);
 247                 }
 248
 249                 SET_KEY_CSUM(k, op->csum);
 250                 if (KEY_CSUM(k))
 251                         bio_csum(n, k);
 252
 253                 trace_bcache_cache_insert(k);
 254                 bch_keylist_push(&op->insert_keys);
 255
 256                 bio_set_op_attrs(n, REQ_OP_WRITE, 0);
 257                 bch_submit_bbio(n, op->c, k, 0);
 258         } while (n != bio);
 259
 260         op->insert_data_done = true;
 261         continue_at(cl, bch_data_insert_keys, op->wq);
 262         return;
 263 err:
 264         /* bch_alloc_sectors() blocks if s->writeback = true */
 265         BUG_ON(op->writeback);
 266
 267         /*
 268          * But if it's not a writeback write we'd rather just bail out if
 269          * there aren't any buckets ready to write to - it might take awhile and
 270          * we might be starving btree writes for gc or something.
 271          */
 272
 273         if (!op->replace) {
 274                 /*
 275                  * Writethrough write: We can't complete the write until we've
 276                  * updated the index. But we don't want to delay the write while
 277                  * we wait for buckets to be freed up, so just invalidate the
 278                  * rest of the write.
 279                  */
 280                 op->bypass = true;
 281                 return bch_data_invalidate(cl);
 282         } else {
 283                 /*
 284                  * From a cache miss, we can just insert the keys for the data
 285                  * we have written or bail out if we didn't do anything.
 286                  */
 287                 op->insert_data_done = true;
 288                 bio_put(bio);
 289
 290                 if (!bch_keylist_empty(&op->insert_keys))
 291                         continue_at(cl, bch_data_insert_keys, op->wq);
 292                 else
 293                         closure_return(cl);
 294         }
 295 }
 296
 297 /**
 298  * bch_data_insert - stick some data in the cache
 299  * @cl: closure pointer.
 300  *
 301  * This is the starting point for any data to end up in a cache device; it could
 302  * be from a normal write, or a writeback write, or a write to a flash only
 303  * volume - it's also used by the moving garbage collector to compact data in
 304  * mostly empty buckets.
 305  *
 306  * It first writes the data to the cache, creating a list of keys to be inserted
 307  * (if the data had to be fragmented there will be multiple keys); after the
 308  * data is written it calls bch_journal, and after the keys have been added to
 309  * the next journal write they're inserted into the btree.
 310  *
 311  * It inserts the data in s->cache_bio; bi_sector is used for the key offset,
 312  * and op->inode is used for the key inode.
 313  *
 314  * If s->bypass is true, instead of inserting the data it invalidates the
 315  * region of the cache represented by s->cache_bio and op->inode.
 316  */
 317 void bch_data_insert(struct closure *cl)
 318 {
 319         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 320
 321         trace_bcache_write(op->c, op->inode, op->bio,
 322                            op->writeback, op->bypass);
 323
 324         bch_keylist_init(&op->insert_keys);
 325         bio_get(op->bio);
 326         bch_data_insert_start(cl);
 327 }
 328
 329 /* Congested? */
 330
 331 unsigned bch_get_congested(struct cache_set *c)
 332 {
 333         int i;
 334         long rand;
 335
 336         if (!c->congested_read_threshold_us &&
 337             !c->congested_write_threshold_us)
 338                 return 0;
 339
 340         i = (local_clock_us() - c->congested_last_us) / 1024;
 341         if (i < 0)
 342                 return 0;
 343
 344         i += atomic_read(&c->congested);
 345         if (i >= 0)
 346                 return 0;
 347
 348         i += CONGESTED_MAX;
 349
 350         if (i > 0)
 351                 i = fract_exp_two(i, 6);
 352
 353         rand = get_random_int();
 354         i -= bitmap_weight(&rand, BITS_PER_LONG);
 355
 356         return i > 0 ? i : 1;
 357 }
 358
 359 static void add_sequential(struct task_struct *t)
 360 {
 361         ewma_add(t->sequential_io_avg,
 362                  t->sequential_io, 8, 0);
 363
 364         t->sequential_io = 0;
 365 }
 366
 367 static struct hlist_head *iohash(struct cached_dev *dc, uint64_t k)
 368 {
 369         return &dc->io_hash[hash_64(k, RECENT_IO_BITS)];
 370 }
 371
 372 static bool check_should_bypass(struct cached_dev *dc, struct bio *bio)
 373 {
 374         struct cache_set *c = dc->disk.c;
 375         unsigned mode = cache_mode(dc);
 376         unsigned sectors, congested = bch_get_congested(c);
 377         struct task_struct *task = current;
 378         struct io *i;
 379
 380         if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
 381             c->gc_stats.in_use > CUTOFF_CACHE_ADD ||
 382             (bio_op(bio) == REQ_OP_DISCARD))
 383                 goto skip;
 384
 385         if (mode == CACHE_MODE_NONE ||
 386             (mode == CACHE_MODE_WRITEAROUND &&
 387              op_is_write(bio_op(bio))))
 388                 goto skip;
 389
 390         /*
 391          * Flag for bypass if the IO is for read-ahead or background,
 392          * unless the read-ahead request is for metadata (eg, for gfs2).
 393          */
 394         if (bio->bi_opf & (REQ_RAHEAD|REQ_BACKGROUND) &&
 395             !(bio->bi_opf & REQ_META))
 396                 goto skip;
 397
 398         if (bio->bi_iter.bi_sector & (c->sb.block_size - 1) ||
 399             bio_sectors(bio) & (c->sb.block_size - 1)) {
 400                 pr_debug("skipping unaligned io");
 401                 goto skip;
 402         }
 403
 404         if (bypass_torture_test(dc)) {
 405                 if ((get_random_int() & 3) == 3)
 406                         goto skip;
 407                 else
 408                         goto rescale;
 409         }
 410
 411         if (!congested && !dc->sequential_cutoff)
 412                 goto rescale;
 413
 414         spin_lock(&dc->io_lock);
 415
 416         hlist_for_each_entry(i, iohash(dc, bio->bi_iter.bi_sector), hash)
 417                 if (i->last == bio->bi_iter.bi_sector &&
 418                     time_before(jiffies, i->jiffies))
 419                         goto found;
 420
 421         i = list_first_entry(&dc->io_lru, struct io, lru);
 422
 423         add_sequential(task);
 424         i->sequential = 0;
 425 found:
 426         if (i->sequential + bio->bi_iter.bi_size > i->sequential)
 427                 i->sequential   += bio->bi_iter.bi_size;
 428
 429         i->last                  = bio_end_sector(bio);
 430         i->jiffies               = jiffies + msecs_to_jiffies(5000);
 431         task->sequential_io      = i->sequential;
 432
 433         hlist_del(&i->hash);
 434         hlist_add_head(&i->hash, iohash(dc, i->last));
 435         list_move_tail(&i->lru, &dc->io_lru);
 436
 437         spin_unlock(&dc->io_lock);
 438
 439         sectors = max(task->sequential_io,
 440                       task->sequential_io_avg) >> 9;
 441
 442         if (dc->sequential_cutoff &&
 443             sectors >= dc->sequential_cutoff >> 9) {
 444                 trace_bcache_bypass_sequential(bio);
 445                 goto skip;
 446         }
 447
 448         if (congested && sectors >= congested) {
 449                 trace_bcache_bypass_congested(bio);
 450                 goto skip;
 451         }
 452
 453 rescale:
 454         bch_rescale_priorities(c, bio_sectors(bio));
 455         return false;
 456 skip:
 457         bch_mark_sectors_bypassed(c, dc, bio_sectors(bio));
 458         return true;
 459 }
 460
 461 /* Cache lookup */
 462
 463 struct search {
 464         /* Stack frame for bio_complete */
 465         struct closure          cl;
 466
 467         struct bbio             bio;
 468         struct bio              *orig_bio;
 469         struct bio              *cache_miss;
 470         struct bcache_device    *d;
 471
 472         unsigned                insert_bio_sectors;
 473         unsigned                recoverable:1;
 474         unsigned                write:1;
 475         unsigned                read_dirty_data:1;
 476         unsigned                cache_missed:1;
 477
 478         unsigned long           start_time;
 479
 480         struct btree_op         op;
 481         struct data_insert_op   iop;
 482 };
 483
 484 static void bch_cache_read_endio(struct bio *bio)
 485 {
 486         struct bbio *b = container_of(bio, struct bbio, bio);
 487         struct closure *cl = bio->bi_private;
 488         struct search *s = container_of(cl, struct search, cl);
 489
 490         /*
 491          * If the bucket was reused while our bio was in flight, we might have
 492          * read the wrong data. Set s->error but not error so it doesn't get
 493          * counted against the cache device, but we'll still reread the data
 494          * from the backing device.
 495          */
 496
 497         if (bio->bi_status)
 498                 s->iop.status = bio->bi_status;
 499         else if (!KEY_DIRTY(&b->key) &&
 500                  ptr_stale(s->iop.c, &b->key, 0)) {
 501                 atomic_long_inc(&s->iop.c->cache_read_races);
 502                 s->iop.status = BLK_STS_IOERR;
 503         }
 504
 505         bch_bbio_endio(s->iop.c, bio, bio->bi_status, "reading from cache");
 506 }
 507
 508 /*
 509  * Read from a single key, handling the initial cache miss if the key starts in
 510  * the middle of the bio
 511  */
 512 static int cache_lookup_fn(struct btree_op *op, struct btree *b, struct bkey *k)
 513 {
 514         struct search *s = container_of(op, struct search, op);
 515         struct bio *n, *bio = &s->bio.bio;
 516         struct bkey *bio_key;
 517         unsigned ptr;
 518
 519         if (bkey_cmp(k, &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0)) <= 0)
 520                 return MAP_CONTINUE;
 521
 522         if (KEY_INODE(k) != s->iop.inode ||
 523             KEY_START(k) > bio->bi_iter.bi_sector) {
 524                 unsigned bio_sectors = bio_sectors(bio);
 525                 unsigned sectors = KEY_INODE(k) == s->iop.inode
 526                         ? min_t(uint64_t, INT_MAX,
 527                                 KEY_START(k) - bio->bi_iter.bi_sector)
 528                         : INT_MAX;
 529
 530                 int ret = s->d->cache_miss(b, s, bio, sectors);
 531                 if (ret != MAP_CONTINUE)
 532                         return ret;
 533
 534                 /* if this was a complete miss we shouldn't get here */
 535                 BUG_ON(bio_sectors <= sectors);
 536         }
 537
 538         if (!KEY_SIZE(k))
 539                 return MAP_CONTINUE;
 540
 541         /* XXX: figure out best pointer - for multiple cache devices */
 542         ptr = 0;
 543
 544         PTR_BUCKET(b->c, k, ptr)->prio = INITIAL_PRIO;
 545
 546         if (KEY_DIRTY(k))
 547                 s->read_dirty_data = true;
 548
 549         n = bio_next_split(bio, min_t(uint64_t, INT_MAX,
 550                                       KEY_OFFSET(k) - bio->bi_iter.bi_sector),
 551                            GFP_NOIO, &s->d->bio_split);
 552
 553         bio_key = &container_of(n, struct bbio, bio)->key;
 554         bch_bkey_copy_single_ptr(bio_key, k, ptr);
 555
 556         bch_cut_front(&KEY(s->iop.inode, n->bi_iter.bi_sector, 0), bio_key);
 557         bch_cut_back(&KEY(s->iop.inode, bio_end_sector(n), 0), bio_key);
 558
 559         n->bi_end_io    = bch_cache_read_endio;
 560         n->bi_private   = &s->cl;
 561
 562         /*
 563          * The bucket we're reading from might be reused while our bio
 564          * is in flight, and we could then end up reading the wrong
 565          * data.
 566          *
 567          * We guard against this by checking (in cache_read_endio()) if
 568          * the pointer is stale again; if so, we treat it as an error
 569          * and reread from the backing device (but we don't pass that
 570          * error up anywhere).
 571          */
 572
 573         __bch_submit_bbio(n, b->c);
 574         return n == bio ? MAP_DONE : MAP_CONTINUE;
 575 }
 576
 577 static void cache_lookup(struct closure *cl)
 578 {
 579         struct search *s = container_of(cl, struct search, iop.cl);
 580         struct bio *bio = &s->bio.bio;
 581         struct cached_dev *dc;
 582         int ret;
 583
 584         bch_btree_op_init(&s->op, -1);
 585
 586         ret = bch_btree_map_keys(&s->op, s->iop.c,
 587                                  &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0),
 588                                  cache_lookup_fn, MAP_END_KEY);
 589         if (ret == -EAGAIN) {
 590                 continue_at(cl, cache_lookup, bcache_wq);
 591                 return;
 592         }
 593
 594         /*
 595          * We might meet err when searching the btree, If that happens, we will
 596          * get negative ret, in this scenario we should not recover data from
 597          * backing device (when cache device is dirty) because we don't know
 598          * whether bkeys the read request covered are all clean.
 599          *
 600          * And after that happened, s->iop.status is still its initial value
 601          * before we submit s->bio.bio
 602          */
 603         if (ret < 0) {
 604                 BUG_ON(ret == -EINTR);
 605                 if (s->d && s->d->c &&
 606                                 !UUID_FLASH_ONLY(&s->d->c->uuids[s->d->id])) {
 607                         dc = container_of(s->d, struct cached_dev, disk);
 608                         if (dc && atomic_read(&dc->has_dirty))
 609                                 s->recoverable = false;
 610                 }
 611                 if (!s->iop.status)
 612                         s->iop.status = BLK_STS_IOERR;
 613         }
 614
 615         closure_return(cl);
 616 }
 617
 618 /* Common code for the make_request functions */
 619
 620 static void request_endio(struct bio *bio)
 621 {
 622         struct closure *cl = bio->bi_private;
 623
 624         if (bio->bi_status) {
 625                 struct search *s = container_of(cl, struct search, cl);
 626                 s->iop.status = bio->bi_status;
 627                 /* Only cache read errors are recoverable */
 628                 s->recoverable = false;
 629         }
 630
 631         bio_put(bio);
 632         closure_put(cl);
 633 }
 634
 635 static void backing_request_endio(struct bio *bio)
 636 {
 637         struct closure *cl = bio->bi_private;
 638
 639         if (bio->bi_status) {
 640                 struct search *s = container_of(cl, struct search, cl);
 641                 struct cached_dev *dc = container_of(s->d,
 642                                                      struct cached_dev, disk);
 643                 /*
 644                  * If a bio has REQ_PREFLUSH for writeback mode, it is
 645                  * speically assembled in cached_dev_write() for a non-zero
 646                  * write request which has REQ_PREFLUSH. we don't set
 647                  * s->iop.status by this failure, the status will be decided
 648                  * by result of bch_data_insert() operation.
 649                  */
 650                 if (unlikely(s->iop.writeback &&
 651                              bio->bi_opf & REQ_PREFLUSH)) {
 652                         pr_err("Can't flush %s: returned bi_status %i",
 653                                 dc->backing_dev_name, bio->bi_status);
 654                 } else {
 655                         /* set to orig_bio->bi_status in bio_complete() */
 656                         s->iop.status = bio->bi_status;
 657                 }
 658                 s->recoverable = false;
 659                 /* should count I/O error for backing device here */
 660                 bch_count_backing_io_errors(dc, bio);
 661         }
 662
 663         bio_put(bio);
 664         closure_put(cl);
 665 }
 666
 667 static void bio_complete(struct search *s)
 668 {
 669         if (s->orig_bio) {
 670                 generic_end_io_acct(s->d->disk->queue,
 671                                     bio_data_dir(s->orig_bio),
 672                                     &s->d->disk->part0, s->start_time);
 673
 674                 trace_bcache_request_end(s->d, s->orig_bio);
 675                 s->orig_bio->bi_status = s->iop.status;
 676                 bio_endio(s->orig_bio);
 677                 s->orig_bio = NULL;
 678         }
 679 }
 680
 681 static void do_bio_hook(struct search *s,
 682                         struct bio *orig_bio,
 683                         bio_end_io_t *end_io_fn)
 684 {
 685         struct bio *bio = &s->bio.bio;
 686
 687         bio_init(bio, NULL, 0);
 688         __bio_clone_fast(bio, orig_bio);
 689         /*
 690          * bi_end_io can be set separately somewhere else, e.g. the
 691          * variants in,
 692          * - cache_bio->bi_end_io from cached_dev_cache_miss()
 693          * - n->bi_end_io from cache_lookup_fn()
 694          */
 695         bio->bi_end_io          = end_io_fn;
 696         bio->bi_private         = &s->cl;
 697
 698         bio_cnt_set(bio, 3);
 699 }
 700
 701 static void search_free(struct closure *cl)
 702 {
 703         struct search *s = container_of(cl, struct search, cl);
 704
 705         if (s->iop.bio)
 706                 bio_put(s->iop.bio);
 707
 708         bio_complete(s);
 709         closure_debug_destroy(cl);
 710         mempool_free(s, &s->d->c->search);
 711 }
 712
 713 static inline struct search *search_alloc(struct bio *bio,
 714                                           struct bcache_device *d)
 715 {
 716         struct search *s;
 717
 718         s = mempool_alloc(&d->c->search, GFP_NOIO);
 719
 720         closure_init(&s->cl, NULL);
 721         do_bio_hook(s, bio, request_endio);
 722
 723         s->orig_bio             = bio;
 724         s->cache_miss           = NULL;
 725         s->cache_missed         = 0;
 726         s->d                    = d;
 727         s->recoverable          = 1;
 728         s->write                = op_is_write(bio_op(bio));
 729         s->read_dirty_data      = 0;
 730         s->start_time           = jiffies;
 731
 732         s->iop.c                = d->c;
 733         s->iop.bio              = NULL;
 734         s->iop.inode            = d->id;
 735         s->iop.write_point      = hash_long((unsigned long) current, 16);
 736         s->iop.write_prio       = 0;
 737         s->iop.status           = 0;
 738         s->iop.flags            = 0;
 739         s->iop.flush_journal    = op_is_flush(bio->bi_opf);
 740         s->iop.wq               = bcache_wq;
 741
 742         return s;
 743 }
 744
 745 /* Cached devices */
 746
 747 static void cached_dev_bio_complete(struct closure *cl)
 748 {
 749         struct search *s = container_of(cl, struct search, cl);
 750         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 751
 752         search_free(cl);
 753         cached_dev_put(dc);
 754 }
 755
 756 /* Process reads */
 757
 758 static void cached_dev_cache_miss_done(struct closure *cl)
 759 {
 760         struct search *s = container_of(cl, struct search, cl);
 761
 762         if (s->iop.replace_collision)
 763                 bch_mark_cache_miss_collision(s->iop.c, s->d);
 764
 765         if (s->iop.bio)
 766                 bio_free_pages(s->iop.bio);
 767
 768         cached_dev_bio_complete(cl);
 769 }
 770
 771 static void cached_dev_read_error(struct closure *cl)
 772 {
 773         struct search *s = container_of(cl, struct search, cl);
 774         struct bio *bio = &s->bio.bio;
 775
 776         /*
 777          * If read request hit dirty data (s->read_dirty_data is true),
 778          * then recovery a failed read request from cached device may
 779          * get a stale data back. So read failure recovery is only
 780          * permitted when read request hit clean data in cache device,
 781          * or when cache read race happened.
 782          */
 783         if (s->recoverable && !s->read_dirty_data) {
 784                 /* Retry from the backing device: */
 785                 trace_bcache_read_retry(s->orig_bio);
 786
 787                 s->iop.status = 0;
 788                 do_bio_hook(s, s->orig_bio, backing_request_endio);
 789
 790                 /* XXX: invalidate cache */
 791
 792                 /* I/O request sent to backing device */
 793                 closure_bio_submit(s->iop.c, bio, cl);
 794         }
 795
 796         continue_at(cl, cached_dev_cache_miss_done, NULL);
 797 }
 798
 799 static void cached_dev_read_done(struct closure *cl)
 800 {
 801         struct search *s = container_of(cl, struct search, cl);
 802         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 803
 804         /*
 805          * We had a cache miss; cache_bio now contains data ready to be inserted
 806          * into the cache.
 807          *
 808          * First, we copy the data we just read from cache_bio's bounce buffers
 809          * to the buffers the original bio pointed to:
 810          */
 811
 812         if (s->iop.bio) {
 813                 bio_reset(s->iop.bio);
 814                 s->iop.bio->bi_iter.bi_sector = s->cache_miss->bi_iter.bi_sector;
 815                 bio_copy_dev(s->iop.bio, s->cache_miss);
 816                 s->iop.bio->bi_iter.bi_size = s->insert_bio_sectors << 9;
 817                 bch_bio_map(s->iop.bio, NULL);
 818
 819                 bio_copy_data(s->cache_miss, s->iop.bio);
 820
 821                 bio_put(s->cache_miss);
 822                 s->cache_miss = NULL;
 823         }
 824
 825         if (verify(dc) && s->recoverable && !s->read_dirty_data)
 826                 bch_data_verify(dc, s->orig_bio);
 827
 828         bio_complete(s);
 829
 830         if (s->iop.bio &&
 831             !test_bit(CACHE_SET_STOPPING, &s->iop.c->flags)) {
 832                 BUG_ON(!s->iop.replace);
 833                 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
 834         }
 835
 836         continue_at(cl, cached_dev_cache_miss_done, NULL);
 837 }
 838
 839 static void cached_dev_read_done_bh(struct closure *cl)
 840 {
 841         struct search *s = container_of(cl, struct search, cl);
 842         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 843
 844         bch_mark_cache_accounting(s->iop.c, s->d,
 845                                   !s->cache_missed, s->iop.bypass);
 846         trace_bcache_read(s->orig_bio, !s->cache_miss, s->iop.bypass);
 847
 848         if (s->iop.status)
 849                 continue_at_nobarrier(cl, cached_dev_read_error, bcache_wq);
 850         else if (s->iop.bio || verify(dc))
 851                 continue_at_nobarrier(cl, cached_dev_read_done, bcache_wq);
 852         else
 853                 continue_at_nobarrier(cl, cached_dev_bio_complete, NULL);
 854 }
 855
 856 static int cached_dev_cache_miss(struct btree *b, struct search *s,
 857                                  struct bio *bio, unsigned sectors)
 858 {
 859         int ret = MAP_CONTINUE;
 860         unsigned reada = 0;
 861         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 862         struct bio *miss, *cache_bio;
 863
 864         s->cache_missed = 1;
 865
 866         if (s->cache_miss || s->iop.bypass) {
 867                 miss = bio_next_split(bio, sectors, GFP_NOIO, &s->d->bio_split);
 868                 ret = miss == bio ? MAP_DONE : MAP_CONTINUE;
 869                 goto out_submit;
 870         }
 871
 872         if (!(bio->bi_opf & REQ_RAHEAD) &&
 873             !(bio->bi_opf & REQ_META) &&
 874             s->iop.c->gc_stats.in_use < CUTOFF_CACHE_READA)
 875                 reada = min_t(sector_t, dc->readahead >> 9,
 876                               get_capacity(bio->bi_disk) - bio_end_sector(bio));
 877
 878         s->insert_bio_sectors = min(sectors, bio_sectors(bio) + reada);
 879
 880         s->iop.replace_key = KEY(s->iop.inode,
 881                                  bio->bi_iter.bi_sector + s->insert_bio_sectors,
 882                                  s->insert_bio_sectors);
 883
 884         ret = bch_btree_insert_check_key(b, &s->op, &s->iop.replace_key);
 885         if (ret)
 886                 return ret;
 887
 888         s->iop.replace = true;
 889
 890         miss = bio_next_split(bio, sectors, GFP_NOIO, &s->d->bio_split);
 891
 892         /* btree_search_recurse()'s btree iterator is no good anymore */
 893         ret = miss == bio ? MAP_DONE : -EINTR;
 894
 895         cache_bio = bio_alloc_bioset(GFP_NOWAIT,
 896                         DIV_ROUND_UP(s->insert_bio_sectors, PAGE_SECTORS),
 897                         &dc->disk.bio_split);
 898         if (!cache_bio)
 899                 goto out_submit;
 900
 901         cache_bio->bi_iter.bi_sector    = miss->bi_iter.bi_sector;
 902         bio_copy_dev(cache_bio, miss);
 903         cache_bio->bi_iter.bi_size      = s->insert_bio_sectors << 9;
 904
 905         cache_bio->bi_end_io    = backing_request_endio;
 906         cache_bio->bi_private   = &s->cl;
 907
 908         bch_bio_map(cache_bio, NULL);
 909         if (bch_bio_alloc_pages(cache_bio, __GFP_NOWARN|GFP_NOIO))
 910                 goto out_put;
 911
 912         if (reada)
 913                 bch_mark_cache_readahead(s->iop.c, s->d);
 914
 915         s->cache_miss   = miss;
 916         s->iop.bio      = cache_bio;
 917         bio_get(cache_bio);
 918         /* I/O request sent to backing device */
 919         closure_bio_submit(s->iop.c, cache_bio, &s->cl);
 920
 921         return ret;
 922 out_put:
 923         bio_put(cache_bio);
 924 out_submit:
 925         miss->bi_end_io         = backing_request_endio;
 926         miss->bi_private        = &s->cl;
 927         /* I/O request sent to backing device */
 928         closure_bio_submit(s->iop.c, miss, &s->cl);
 929         return ret;
 930 }
 931
 932 static void cached_dev_read(struct cached_dev *dc, struct search *s)
 933 {
 934         struct closure *cl = &s->cl;
 935
 936         closure_call(&s->iop.cl, cache_lookup, NULL, cl);
 937         continue_at(cl, cached_dev_read_done_bh, NULL);
 938 }
 939
 940 /* Process writes */
 941
 942 static void cached_dev_write_complete(struct closure *cl)
 943 {
 944         struct search *s = container_of(cl, struct search, cl);
 945         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 946
 947         up_read_non_owner(&dc->writeback_lock);
 948         cached_dev_bio_complete(cl);
 949 }
 950
 951 static void cached_dev_write(struct cached_dev *dc, struct search *s)
 952 {
 953         struct closure *cl = &s->cl;
 954         struct bio *bio = &s->bio.bio;
 955         struct bkey start = KEY(dc->disk.id, bio->bi_iter.bi_sector, 0);
 956         struct bkey end = KEY(dc->disk.id, bio_end_sector(bio), 0);
 957
 958         bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys, &start, &end);
 959
 960         down_read_non_owner(&dc->writeback_lock);
 961         if (bch_keybuf_check_overlapping(&dc->writeback_keys, &start, &end)) {
 962                 /*
 963                  * We overlap with some dirty data undergoing background
 964                  * writeback, force this write to writeback
 965                  */
 966                 s->iop.bypass = false;
 967                 s->iop.writeback = true;
 968         }
 969
 970         /*
 971          * Discards aren't _required_ to do anything, so skipping if
 972          * check_overlapping returned true is ok
 973          *
 974          * But check_overlapping drops dirty keys for which io hasn't started,
 975          * so we still want to call it.
 976          */
 977         if (bio_op(bio) == REQ_OP_DISCARD)
 978                 s->iop.bypass = true;
 979
 980         if (should_writeback(dc, s->orig_bio,
 981                              cache_mode(dc),
 982                              s->iop.bypass)) {
 983                 s->iop.bypass = false;
 984                 s->iop.writeback = true;
 985         }
 986
 987         if (s->iop.bypass) {
 988                 s->iop.bio = s->orig_bio;
 989                 bio_get(s->iop.bio);
 990
 991                 if (bio_op(bio) == REQ_OP_DISCARD &&
 992                     !blk_queue_discard(bdev_get_queue(dc->bdev)))
 993                         goto insert_data;
 994
 995                 /* I/O request sent to backing device */
 996                 bio->bi_end_io = backing_request_endio;
 997                 closure_bio_submit(s->iop.c, bio, cl);
 998
 999         } else if (s->iop.writeback) {
1000                 bch_writeback_add(dc);
1001                 s->iop.bio = bio;
1002
1003                 if (bio->bi_opf & REQ_PREFLUSH) {
1004                         /*
1005                          * Also need to send a flush to the backing
1006                          * device.
1007                          */
1008                         struct bio *flush;
1009
1010                         flush = bio_alloc_bioset(GFP_NOIO, 0,
1011                                                  &dc->disk.bio_split);
1012                         if (!flush) {
1013                                 s->iop.status = BLK_STS_RESOURCE;
1014                                 goto insert_data;
1015                         }
1016                         bio_copy_dev(flush, bio);
1017                         flush->bi_end_io = backing_request_endio;
1018                         flush->bi_private = cl;
1019                         flush->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
1020                         /* I/O request sent to backing device */
1021                         closure_bio_submit(s->iop.c, flush, cl);
1022                 }
1023         } else {
1024                 s->iop.bio = bio_clone_fast(bio, GFP_NOIO, &dc->disk.bio_split);
1025                 /* I/O request sent to backing device */
1026                 bio->bi_end_io = backing_request_endio;
1027                 closure_bio_submit(s->iop.c, bio, cl);
1028         }
1029
1030 insert_data:
1031         closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
1032         continue_at(cl, cached_dev_write_complete, NULL);
1033 }
1034
1035 static void cached_dev_nodata(struct closure *cl)
1036 {
1037         struct search *s = container_of(cl, struct search, cl);
1038         struct bio *bio = &s->bio.bio;
1039
1040         if (s->iop.flush_journal)
1041                 bch_journal_meta(s->iop.c, cl);
1042
1043         /* If it's a flush, we send the flush to the backing device too */
1044         bio->bi_end_io = backing_request_endio;
1045         closure_bio_submit(s->iop.c, bio, cl);
1046
1047         continue_at(cl, cached_dev_bio_complete, NULL);
1048 }
1049
1050 struct detached_dev_io_private {
1051         struct bcache_device    *d;
1052         unsigned long           start_time;
1053         bio_end_io_t            *bi_end_io;
1054         void                    *bi_private;
1055 };
1056
1057 static void detached_dev_end_io(struct bio *bio)
1058 {
1059         struct detached_dev_io_private *ddip;
1060
1061         ddip = bio->bi_private;
1062         bio->bi_end_io = ddip->bi_end_io;
1063         bio->bi_private = ddip->bi_private;
1064
1065         generic_end_io_acct(ddip->d->disk->queue,
1066                             bio_data_dir(bio),
1067                             &ddip->d->disk->part0, ddip->start_time);
1068
1069         if (bio->bi_status) {
1070                 struct cached_dev *dc = container_of(ddip->d,
1071                                                      struct cached_dev, disk);
1072                 /* should count I/O error for backing device here */
1073                 bch_count_backing_io_errors(dc, bio);
1074         }
1075
1076         kfree(ddip);
1077         bio->bi_end_io(bio);
1078 }
1079
1080 static void detached_dev_do_request(struct bcache_device *d, struct bio *bio)
1081 {
1082         struct detached_dev_io_private *ddip;
1083         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1084
1085         /*
1086          * no need to call closure_get(&dc->disk.cl),
1087          * because upper layer had already opened bcache device,
1088          * which would call closure_get(&dc->disk.cl)
1089          */
1090         ddip = kzalloc(sizeof(struct detached_dev_io_private), GFP_NOIO);
1091         ddip->d = d;
1092         ddip->start_time = jiffies;
1093         ddip->bi_end_io = bio->bi_end_io;
1094         ddip->bi_private = bio->bi_private;
1095         bio->bi_end_io = detached_dev_end_io;
1096         bio->bi_private = ddip;
1097
1098         if ((bio_op(bio) == REQ_OP_DISCARD) &&
1099             !blk_queue_discard(bdev_get_queue(dc->bdev)))
1100                 bio->bi_end_io(bio);
1101         else
1102                 generic_make_request(bio);
1103 }
1104
1105 /* Cached devices - read & write stuff */
1106
1107 static blk_qc_t cached_dev_make_request(struct request_queue *q,
1108                                         struct bio *bio)
1109 {
1110         struct search *s;
1111         struct bcache_device *d = bio->bi_disk->private_data;
1112         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1113         int rw = bio_data_dir(bio);
1114
1115         if (unlikely((d->c && test_bit(CACHE_SET_IO_DISABLE, &d->c->flags)) ||
1116                      dc->io_disable)) {
1117                 bio->bi_status = BLK_STS_IOERR;
1118                 bio_endio(bio);
1119                 return BLK_QC_T_NONE;
1120         }
1121
1122         atomic_set(&dc->backing_idle, 0);
1123         generic_start_io_acct(q, rw, bio_sectors(bio), &d->disk->part0);
1124
1125         bio_set_dev(bio, dc->bdev);
1126         bio->bi_iter.bi_sector += dc->sb.data_offset;
1127
1128         if (cached_dev_get(dc)) {
1129                 s = search_alloc(bio, d);
1130                 trace_bcache_request_start(s->d, bio);
1131
1132                 if (!bio->bi_iter.bi_size) {
1133                         /*
1134                          * can't call bch_journal_meta from under
1135                          * generic_make_request
1136                          */
1137                         continue_at_nobarrier(&s->cl,
1138                                               cached_dev_nodata,
1139                                               bcache_wq);
1140                 } else {
1141                         s->iop.bypass = check_should_bypass(dc, bio);
1142
1143                         if (rw)
1144                                 cached_dev_write(dc, s);
1145                         else
1146                                 cached_dev_read(dc, s);
1147                 }
1148         } else
1149                 /* I/O request sent to backing device */
1150                 detached_dev_do_request(d, bio);
1151
1152         return BLK_QC_T_NONE;
1153 }
1154
1155 static int cached_dev_ioctl(struct bcache_device *d, fmode_t mode,
1156                             unsigned int cmd, unsigned long arg)
1157 {
1158         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1159         return __blkdev_driver_ioctl(dc->bdev, mode, cmd, arg);
1160 }
1161
1162 static int cached_dev_congested(void *data, int bits)
1163 {
1164         struct bcache_device *d = data;
1165         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1166         struct request_queue *q = bdev_get_queue(dc->bdev);
1167         int ret = 0;
1168
1169         if (bdi_congested(q->backing_dev_info, bits))
1170                 return 1;
1171
1172         if (cached_dev_get(dc)) {
1173                 unsigned i;
1174                 struct cache *ca;
1175
1176                 for_each_cache(ca, d->c, i) {
1177                         q = bdev_get_queue(ca->bdev);
1178                         ret |= bdi_congested(q->backing_dev_info, bits);
1179                 }
1180
1181                 cached_dev_put(dc);
1182         }
1183
1184         return ret;
1185 }
1186
1187 void bch_cached_dev_request_init(struct cached_dev *dc)
1188 {
1189         struct gendisk *g = dc->disk.disk;
1190
1191         g->queue->make_request_fn               = cached_dev_make_request;
1192         g->queue->backing_dev_info->congested_fn = cached_dev_congested;
1193         dc->disk.cache_miss                     = cached_dev_cache_miss;
1194         dc->disk.ioctl                          = cached_dev_ioctl;
1195 }
1196
1197 /* Flash backed devices */
1198
1199 static int flash_dev_cache_miss(struct btree *b, struct search *s,
1200                                 struct bio *bio, unsigned sectors)
1201 {
1202         unsigned bytes = min(sectors, bio_sectors(bio)) << 9;
1203
1204         swap(bio->bi_iter.bi_size, bytes);
1205         zero_fill_bio(bio);
1206         swap(bio->bi_iter.bi_size, bytes);
1207
1208         bio_advance(bio, bytes);
1209
1210         if (!bio->bi_iter.bi_size)
1211                 return MAP_DONE;
1212
1213         return MAP_CONTINUE;
1214 }
1215
1216 static void flash_dev_nodata(struct closure *cl)
1217 {
1218         struct search *s = container_of(cl, struct search, cl);
1219
1220         if (s->iop.flush_journal)
1221                 bch_journal_meta(s->iop.c, cl);
1222
1223         continue_at(cl, search_free, NULL);
1224 }
1225
1226 static blk_qc_t flash_dev_make_request(struct request_queue *q,
1227                                              struct bio *bio)
1228 {
1229         struct search *s;
1230         struct closure *cl;
1231         struct bcache_device *d = bio->bi_disk->private_data;
1232         int rw = bio_data_dir(bio);
1233
1234         if (unlikely(d->c && test_bit(CACHE_SET_IO_DISABLE, &d->c->flags))) {
1235                 bio->bi_status = BLK_STS_IOERR;
1236                 bio_endio(bio);
1237                 return BLK_QC_T_NONE;
1238         }
1239
1240         generic_start_io_acct(q, rw, bio_sectors(bio), &d->disk->part0);
1241
1242         s = search_alloc(bio, d);
1243         cl = &s->cl;
1244         bio = &s->bio.bio;
1245
1246         trace_bcache_request_start(s->d, bio);
1247
1248         if (!bio->bi_iter.bi_size) {
1249                 /*
1250                  * can't call bch_journal_meta from under
1251                  * generic_make_request
1252                  */
1253                 continue_at_nobarrier(&s->cl,
1254                                       flash_dev_nodata,
1255                                       bcache_wq);
1256                 return BLK_QC_T_NONE;
1257         } else if (rw) {
1258                 bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys,
1259                                         &KEY(d->id, bio->bi_iter.bi_sector, 0),
1260                                         &KEY(d->id, bio_end_sector(bio), 0));
1261
1262                 s->iop.bypass           = (bio_op(bio) == REQ_OP_DISCARD) != 0;
1263                 s->iop.writeback        = true;
1264                 s->iop.bio              = bio;
1265
1266                 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
1267         } else {
1268                 closure_call(&s->iop.cl, cache_lookup, NULL, cl);
1269         }
1270
1271         continue_at(cl, search_free, NULL);
1272         return BLK_QC_T_NONE;
1273 }
1274
1275 static int flash_dev_ioctl(struct bcache_device *d, fmode_t mode,
1276                            unsigned int cmd, unsigned long arg)
1277 {
1278         return -ENOTTY;
1279 }
1280
1281 static int flash_dev_congested(void *data, int bits)
1282 {
1283         struct bcache_device *d = data;
1284         struct request_queue *q;
1285         struct cache *ca;
1286         unsigned i;
1287         int ret = 0;
1288
1289         for_each_cache(ca, d->c, i) {
1290                 q = bdev_get_queue(ca->bdev);
1291                 ret |= bdi_congested(q->backing_dev_info, bits);
1292         }
1293
1294         return ret;
1295 }
1296
1297 void bch_flash_dev_request_init(struct bcache_device *d)
1298 {
1299         struct gendisk *g = d->disk;
1300
1301         g->queue->make_request_fn               = flash_dev_make_request;
1302         g->queue->backing_dev_info->congested_fn = flash_dev_congested;
1303         d->cache_miss                           = flash_dev_cache_miss;
1304         d->ioctl                                = flash_dev_ioctl;
1305 }
1306
1307 void bch_request_exit(void)
1308 {
1309         if (bch_search_cache)
1310                 kmem_cache_destroy(bch_search_cache);
1311 }
1312
1313 int __init bch_request_init(void)
1314 {
1315         bch_search_cache = KMEM_CACHE(search, 0);
1316         if (!bch_search_cache)
1317                 return -ENOMEM;
1318
1319         return 0;
1320 }